1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2025-01-31 20:51:52 +01:00

AArch64: Cyclone: Remove SlowMisaligned128Store tuning flag

Remove FeatureSlowMisaligned128Store from cyclone flags.
This flag causes splitting of 16 byte wide stores into 2 stored of 8
bytes. This was useful on older apple CPUs which were slow for 16byte
stores that were not aligned on 16byte. As the compiler often cannot
predict the actual alignment, the splitting was choosen.

This has been a topic for a lot of debate as the splitting also
decreases performance for some benchmarks. Measuring the effects on
newer apple chips (rdar://35525421) shows that it harms more cases than
it helps. So it is time to retire this workaround.

llvm-svn: 323289
This commit is contained in:
Matthias Braun 2018-01-24 00:39:53 +00:00
parent c6aca8a06d
commit 81d794ab6d
5 changed files with 12 additions and 15 deletions

View File

@ -294,7 +294,6 @@ def ProcCyclone : SubtargetFeature<"cyclone", "ARMProcFamily", "Cyclone",
FeatureFuseAES,
FeatureNEON,
FeaturePerfMon,
FeatureSlowMisaligned128Store,
FeatureZCRegMove,
FeatureZCZeroing,
FeatureZCZeroingFPWorkaround

View File

@ -105,8 +105,8 @@ define void @splat_v4i32(i32 %v, i32 *%p) {
entry:
; CHECK-LABEL: splat_v4i32
; CHECK-DAG: stp w0, w0, [x1]
; CHECK-DAG: stp w0, w0, [x1, #8]
; CHECK-DAG: dup v0.4s, w0
; CHECK-DAG: str q0, [x1]
; CHECK: ret
%p17 = insertelement <4 x i32> undef, i32 %v, i32 0
@ -129,8 +129,7 @@ entry:
; CHECK-DAG: mov v[[REG1]].s[1], w0
; CHECK-DAG: mov v[[REG1]].s[2], w0
; CHECK-DAG: mov v[[REG1]].s[3], w0
; CHECK: ext v[[REG2:[0-9]+]].16b, v[[REG1]].16b, v[[REG1]].16b, #8
; CHECK: stp d[[REG1]], d[[REG2]], [x1]
; CHECK: str q[[REG1]], [x1]
; CHECK: ret
%p17 = insertelement <4 x i32> undef, i32 %v, i32 %v
@ -151,8 +150,7 @@ entry:
; CHECK: mov v[[REG1]].s[1], w0
; CHECK-DAG: mov v[[REG1]].s[2], w0
; CHECK-DAG: mov v[[REG1]].s[3], w0
; CHECK: ext v[[REG2:[0-9]+]].16b, v[[REG1]].16b, v[[REG1]].16b, #8
; CHECK: stp d[[REG1]], d[[REG2]], [x1]
; CHECK: str q[[REG1]], [x1]
; CHECK: ret
%p18 = insertelement <4 x i32> %vin, i32 %v, i32 1

View File

@ -1,4 +1,4 @@
; RUN: llc < %s -mtriple=arm64-eabi -aarch64-neon-syntax=apple -mcpu=cyclone | FileCheck %s
; RUN: llc < %s -mtriple=arm64-eabi -aarch64-neon-syntax=apple -mcpu=cyclone -mattr=+slow-misaligned-128store | FileCheck %s
%struct.X = type <{ i32, i64, i64 }>
define void @foo1(i32* %p, i64 %val) nounwind {

View File

@ -1,4 +1,4 @@
; RUN: llc < %s -mtriple=aarch64-unknown-unknown -mcpu=cyclone | FileCheck %s --check-prefix=CYCLONE --check-prefix=CHECK
; RUN: llc < %s -mtriple=aarch64-unknown-unknown -mcpu=cyclone -mattr=+slow-misaligned-128store | FileCheck %s --check-prefix=SPLITTING --check-prefix=CHECK
; RUN: llc < %s -mtriple=aarch64-eabi -mattr=-slow-misaligned-128store | FileCheck %s --check-prefix=MISALIGNED --check-prefix=CHECK
@g0 = external global <3 x float>, align 16
@ -44,9 +44,9 @@ define void @merge_vec_extract_stores(<4 x float> %v1, <2 x float>* %ptr) {
; FIXME: Ideally we would like to use a generic target for this test, but this relies
; on suppressing store pairs.
; CYCLONE-LABEL: merge_vec_extract_stores
; CYCLONE: ext v1.16b, v0.16b, v0.16b, #8
; CYCLONE-NEXT: str d0, [x0, #24]
; CYCLONE-NEXT: str d1, [x0, #32]
; CYCLONE-NEXT: ret
; SPLITTING-LABEL: merge_vec_extract_stores
; SPLITTING: ext v1.16b, v0.16b, v0.16b, #8
; SPLITTING-NEXT: str d0, [x0, #24]
; SPLITTING-NEXT: str d1, [x0, #32]
; SPLITTING-NEXT: ret
}

View File

@ -1,5 +1,5 @@
; REQUIRES: asserts
; RUN: llc < %s -mtriple=aarch64 -mcpu=cyclone -mattr=+use-aa -enable-misched -verify-misched -o - | FileCheck %s
; RUN: llc < %s -mtriple=aarch64 -mcpu=cyclone -mattr=+use-aa,+slow-misaligned-128store -enable-misched -verify-misched -o - | FileCheck %s
; Tests to check that the scheduler dependencies derived from alias analysis are
; correct when we have loads that have been split up so that they can later be